mm/page-writeback.c

   1 /*
   2  * mm/page-writeback.c.
   3  *
   4  * Copyright (C) 2002, Linus Torvalds.
   5  *
   6  * Contains functions related to writing back dirty pages at the
   7  * address_space level.
   8  *
   9  * 10Apr2002    akpm@zip.com.au
  10  *              Initial version
  11  */
  12
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/fs.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/slab.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/init.h>
  23 #include <linux/backing-dev.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/mpage.h>
  26 #include <linux/rmap.h>
  27 #include <linux/percpu.h>
  28 #include <linux/notifier.h>
  29 #include <linux/smp.h>
  30 #include <linux/sysctl.h>
  31 #include <linux/cpu.h>
  32 #include <linux/syscalls.h>
  33
  34 /*
  35  * The maximum number of pages to writeout in a single bdflush/kupdate
  36  * operation.  We do this so we don't hold I_LOCK against an inode for
  37  * enormous amounts of time, which would block a userspace task which has
  38  * been forced to throttle against that inode.  Also, the code reevaluates
  39  * the dirty each time it has written this many pages.
  40  */
  41 #define MAX_WRITEBACK_PAGES     1024
  42
  43 /*
  44  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  45  * will look to see if it needs to force writeback or throttling.
  46  */
  47 static long ratelimit_pages = 32;
  48
  49 static long total_pages;        /* The total number of pages in the machine. */
  50 static int dirty_exceeded __cacheline_aligned_in_smp;   /* Dirty mem may be over limit */
  51
  52 /*
  53  * When balance_dirty_pages decides that the caller needs to perform some
  54  * non-background writeback, this is how many pages it will attempt to write.
  55  * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
  56  * large amounts of I/O are submitted.
  57  */
  58 static inline long sync_writeback_pages(void)
  59 {
  60         return ratelimit_pages + ratelimit_pages / 2;
  61 }
  62
  63 /* The following parameters are exported via /proc/sys/vm */
  64
  65 /*
  66  * Start background writeback (via pdflush) at this percentage
  67  */
  68 int dirty_background_ratio = 10;
  69
  70 /*
  71  * The generator of dirty data starts writeback at this percentage
  72  */
  73 int vm_dirty_ratio = 40;
  74
  75 /*
  76  * The interval between `kupdate'-style writebacks, in jiffies
  77  */
  78 int dirty_writeback_interval = 5 * HZ;
  79
  80 /*
  81  * The longest number of jiffies for which data is allowed to remain dirty
  82  */
  83 int dirty_expire_interval = 30 * HZ;
  84
  85 /*
  86  * Flag that makes the machine dump writes/reads and block dirtyings.
  87  */
  88 int block_dump;
  89
  90 /*
  91  * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
  92  * a full sync is triggered after this time elapses without any disk activity.
  93  */
  94 int laptop_mode;
  95
  96 EXPORT_SYMBOL(laptop_mode);
  97
  98 /* End of sysctl-exported parameters */
  99
 100
 101 static void background_writeout(unsigned long _min_pages);
 102
 103 /*
 104  * Work out the current dirty-memory clamping and background writeout
 105  * thresholds.
 106  *
 107  * The main aim here is to lower them aggressively if there is a lot of mapped
 108  * memory around.  To avoid stressing page reclaim with lots of unreclaimable
 109  * pages.  It is better to clamp down on writers than to start swapping, and
 110  * performing lots of scanning.
 111  *
 112  * We only allow 1/2 of the currently-unmapped memory to be dirtied.
 113  *
 114  * We don't permit the clamping level to fall below 5% - that is getting rather
 115  * excessive.
 116  *
 117  * We make sure that the background writeout level is below the adjusted
 118  * clamping level.
 119  */
 120 static void
 121 get_dirty_limits(long *pbackground, long *pdirty,
 122                                         struct address_space *mapping)
 123 {
 124         int background_ratio;           /* Percentages */
 125         int dirty_ratio;
 126         int unmapped_ratio;
 127         long background;
 128         long dirty;
 129         unsigned long available_memory = total_pages;
 130         struct task_struct *tsk;
 131
 132 #ifdef CONFIG_HIGHMEM
 133         /*
 134          * If this mapping can only allocate from low memory,
 135          * we exclude high memory from our count.
 136          */
 137         if (mapping && !(mapping_gfp_mask(mapping) & __GFP_HIGHMEM))
 138                 available_memory -= totalhigh_pages;
 139 #endif
 140
 141
 142         unmapped_ratio = 100 - ((global_page_state(NR_FILE_MAPPED) +
 143                                 global_page_state(NR_ANON_PAGES)) * 100) /
 144                                         total_pages;
 145
 146         dirty_ratio = vm_dirty_ratio;
 147         if (dirty_ratio > unmapped_ratio / 2)
 148                 dirty_ratio = unmapped_ratio / 2;
 149
 150         if (dirty_ratio < 5)
 151                 dirty_ratio = 5;
 152
 153         background_ratio = dirty_background_ratio;
 154         if (background_ratio >= dirty_ratio)
 155                 background_ratio = dirty_ratio / 2;
 156
 157         background = (background_ratio * available_memory) / 100;
 158         dirty = (dirty_ratio * available_memory) / 100;
 159         tsk = current;
 160         if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
 161                 background += background / 4;
 162                 dirty += dirty / 4;
 163         }
 164         *pbackground = background;
 165         *pdirty = dirty;
 166 }
 167
 168 /*
 169  * balance_dirty_pages() must be called by processes which are generating dirty
 170  * data.  It looks at the number of dirty pages in the machine and will force
 171  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
 172  * If we're over `background_thresh' then pdflush is woken to perform some
 173  * writeout.
 174  */
 175 static void balance_dirty_pages(struct address_space *mapping)
 176 {
 177         long nr_reclaimable;
 178         long background_thresh;
 179         long dirty_thresh;
 180         unsigned long pages_written = 0;
 181         unsigned long write_chunk = sync_writeback_pages();
 182
 183         struct backing_dev_info *bdi = mapping->backing_dev_info;
 184
 185         for (;;) {
 186                 struct writeback_control wbc = {
 187                         .bdi            = bdi,
 188                         .sync_mode      = WB_SYNC_NONE,
 189                         .older_than_this = NULL,
 190                         .nr_to_write    = write_chunk,
 191                         .range_cyclic   = 1,
 192                 };
 193
 194                 get_dirty_limits(&background_thresh, &dirty_thresh, mapping);
 195                 nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
 196                                         global_page_state(NR_UNSTABLE_NFS);
 197                 if (nr_reclaimable + global_page_state(NR_WRITEBACK) <=
 198                         dirty_thresh)
 199                                 break;
 200
 201                 if (!dirty_exceeded)
 202                         dirty_exceeded = 1;
 203
 204                 /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
 205                  * Unstable writes are a feature of certain networked
 206                  * filesystems (i.e. NFS) in which data may have been
 207                  * written to the server's write cache, but has not yet
 208                  * been flushed to permanent storage.
 209                  */
 210                 if (nr_reclaimable) {
 211                         writeback_inodes(&wbc);
 212                         get_dirty_limits(&background_thresh,
 213                                                 &dirty_thresh, mapping);
 214                         nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
 215                                         global_page_state(NR_UNSTABLE_NFS);
 216                         if (nr_reclaimable +
 217                                 global_page_state(NR_WRITEBACK)
 218                                         <= dirty_thresh)
 219                                                 break;
 220                         pages_written += write_chunk - wbc.nr_to_write;
 221                         if (pages_written >= write_chunk)
 222                                 break;          /* We've done our duty */
 223                 }
 224                 blk_congestion_wait(WRITE, HZ/10);
 225         }
 226
 227         if (nr_reclaimable + global_page_state(NR_WRITEBACK)
 228                 <= dirty_thresh && dirty_exceeded)
 229                         dirty_exceeded = 0;
 230
 231         if (writeback_in_progress(bdi))
 232                 return;         /* pdflush is already working this queue */
 233
 234         /*
 235          * In laptop mode, we wait until hitting the higher threshold before
 236          * starting background writeout, and then write out all the way down
 237          * to the lower threshold.  So slow writers cause minimal disk activity.
 238          *
 239          * In normal mode, we start background writeout at the lower
 240          * background_thresh, to keep the amount of dirty memory low.
 241          */
 242         if ((laptop_mode && pages_written) ||
 243              (!laptop_mode && (nr_reclaimable > background_thresh)))
 244                 pdflush_operation(background_writeout, 0);
 245 }
 246
 247 /**
 248  * balance_dirty_pages_ratelimited_nr - balance dirty memory state
 249  * @mapping: address_space which was dirtied
 250  * @nr_pages_dirtied: number of pages which the caller has just dirtied
 251  *
 252  * Processes which are dirtying memory should call in here once for each page
 253  * which was newly dirtied.  The function will periodically check the system's
 254  * dirty state and will initiate writeback if needed.
 255  *
 256  * On really big machines, get_writeback_state is expensive, so try to avoid
 257  * calling it too often (ratelimiting).  But once we're over the dirty memory
 258  * limit we decrease the ratelimiting by a lot, to prevent individual processes
 259  * from overshooting the limit by (ratelimit_pages) each.
 260  */
 261 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
 262                                         unsigned long nr_pages_dirtied)
 263 {
 264         static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
 265         unsigned long ratelimit;
 266         unsigned long *p;
 267
 268         ratelimit = ratelimit_pages;
 269         if (dirty_exceeded)
 270                 ratelimit = 8;
 271
 272         /*
 273          * Check the rate limiting. Also, we do not want to throttle real-time
 274          * tasks in balance_dirty_pages(). Period.
 275          */
 276         preempt_disable();
 277         p =  &__get_cpu_var(ratelimits);
 278         *p += nr_pages_dirtied;
 279         if (unlikely(*p >= ratelimit)) {
 280                 *p = 0;
 281                 preempt_enable();
 282                 balance_dirty_pages(mapping);
 283                 return;
 284         }
 285         preempt_enable();
 286 }
 287 EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
 288
 289 void throttle_vm_writeout(void)
 290 {
 291         long background_thresh;
 292         long dirty_thresh;
 293
 294         for ( ; ; ) {
 295                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
 296
 297                 /*
 298                  * Boost the allowable dirty threshold a bit for page
 299                  * allocators so they don't get DoS'ed by heavy writers
 300                  */
 301                 dirty_thresh += dirty_thresh / 10;      /* wheeee... */
 302
 303                 if (global_page_state(NR_UNSTABLE_NFS) +
 304                         global_page_state(NR_WRITEBACK) <= dirty_thresh)
 305                                 break;
 306                 blk_congestion_wait(WRITE, HZ/10);
 307         }
 308 }
 309
 310
 311 /*
 312  * writeback at least _min_pages, and keep writing until the amount of dirty
 313  * memory is less than the background threshold, or until we're all clean.
 314  */
 315 static void background_writeout(unsigned long _min_pages)
 316 {
 317         long min_pages = _min_pages;
 318         struct writeback_control wbc = {
 319                 .bdi            = NULL,
 320                 .sync_mode      = WB_SYNC_NONE,
 321                 .older_than_this = NULL,
 322                 .nr_to_write    = 0,
 323                 .nonblocking    = 1,
 324                 .range_cyclic   = 1,
 325         };
 326
 327         for ( ; ; ) {
 328                 long background_thresh;
 329                 long dirty_thresh;
 330
 331                 get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
 332                 if (global_page_state(NR_FILE_DIRTY) +
 333                         global_page_state(NR_UNSTABLE_NFS) < background_thresh
 334                                 && min_pages <= 0)
 335                         break;
 336                 wbc.encountered_congestion = 0;
 337                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 338                 wbc.pages_skipped = 0;
 339                 writeback_inodes(&wbc);
 340                 min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 341                 if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
 342                         /* Wrote less than expected */
 343                         blk_congestion_wait(WRITE, HZ/10);
 344                         if (!wbc.encountered_congestion)
 345                                 break;
 346                 }
 347         }
 348 }
 349
 350 /*
 351  * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
 352  * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
 353  * -1 if all pdflush threads were busy.
 354  */
 355 int wakeup_pdflush(long nr_pages)
 356 {
 357         if (nr_pages == 0)
 358                 nr_pages = global_page_state(NR_FILE_DIRTY) +
 359                                 global_page_state(NR_UNSTABLE_NFS);
 360         return pdflush_operation(background_writeout, nr_pages);
 361 }
 362
 363 static void wb_timer_fn(unsigned long unused);
 364 static void laptop_timer_fn(unsigned long unused);
 365
 366 static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
 367 static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
 368
 369 /*
 370  * Periodic writeback of "old" data.
 371  *
 372  * Define "old": the first time one of an inode's pages is dirtied, we mark the
 373  * dirtying-time in the inode's address_space.  So this periodic writeback code
 374  * just walks the superblock inode list, writing back any inodes which are
 375  * older than a specific point in time.
 376  *
 377  * Try to run once per dirty_writeback_interval.  But if a writeback event
 378  * takes longer than a dirty_writeback_interval interval, then leave a
 379  * one-second gap.
 380  *
 381  * older_than_this takes precedence over nr_to_write.  So we'll only write back
 382  * all dirty pages if they are all attached to "old" mappings.
 383  */
 384 static void wb_kupdate(unsigned long arg)
 385 {
 386         unsigned long oldest_jif;
 387         unsigned long start_jif;
 388         unsigned long next_jif;
 389         long nr_to_write;
 390         struct writeback_control wbc = {
 391                 .bdi            = NULL,
 392                 .sync_mode      = WB_SYNC_NONE,
 393                 .older_than_this = &oldest_jif,
 394                 .nr_to_write    = 0,
 395                 .nonblocking    = 1,
 396                 .for_kupdate    = 1,
 397                 .range_cyclic   = 1,
 398         };
 399
 400         sync_supers();
 401
 402         oldest_jif = jiffies - dirty_expire_interval;
 403         start_jif = jiffies;
 404         next_jif = start_jif + dirty_writeback_interval;
 405         nr_to_write = global_page_state(NR_FILE_DIRTY) +
 406                         global_page_state(NR_UNSTABLE_NFS) +
 407                         (inodes_stat.nr_inodes - inodes_stat.nr_unused);
 408         while (nr_to_write > 0) {
 409                 wbc.encountered_congestion = 0;
 410                 wbc.nr_to_write = MAX_WRITEBACK_PAGES;
 411                 writeback_inodes(&wbc);
 412                 if (wbc.nr_to_write > 0) {
 413                         if (wbc.encountered_congestion)
 414                                 blk_congestion_wait(WRITE, HZ/10);
 415                         else
 416                                 break;  /* All the old data is written */
 417                 }
 418                 nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 419         }
 420         if (time_before(next_jif, jiffies + HZ))
 421                 next_jif = jiffies + HZ;
 422         if (dirty_writeback_interval)
 423                 mod_timer(&wb_timer, next_jif);
 424 }
 425
 426 /*
 427  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
 428  */
 429 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
 430                 struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
 431 {
 432         proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
 433         if (dirty_writeback_interval) {
 434                 mod_timer(&wb_timer,
 435                         jiffies + dirty_writeback_interval);
 436                 } else {
 437                 del_timer(&wb_timer);
 438         }
 439         return 0;
 440 }
 441
 442 static void wb_timer_fn(unsigned long unused)
 443 {
 444         if (pdflush_operation(wb_kupdate, 0) < 0)
 445                 mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
 446 }
 447
 448 static void laptop_flush(unsigned long unused)
 449 {
 450         sys_sync();
 451 }
 452
 453 static void laptop_timer_fn(unsigned long unused)
 454 {
 455         pdflush_operation(laptop_flush, 0);
 456 }
 457
 458 /*
 459  * We've spun up the disk and we're in laptop mode: schedule writeback
 460  * of all dirty data a few seconds from now.  If the flush is already scheduled
 461  * then push it back - the user is still using the disk.
 462  */
 463 void laptop_io_completion(void)
 464 {
 465         mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
 466 }
 467
 468 /*
 469  * We're in laptop mode and we've just synced. The sync's writes will have
 470  * caused another writeback to be scheduled by laptop_io_completion.
 471  * Nothing needs to be written back anymore, so we unschedule the writeback.
 472  */
 473 void laptop_sync_completion(void)
 474 {
 475         del_timer(&laptop_mode_wb_timer);
 476 }
 477
 478 /*
 479  * If ratelimit_pages is too high then we can get into dirty-data overload
 480  * if a large number of processes all perform writes at the same time.
 481  * If it is too low then SMP machines will call the (expensive)
 482  * get_writeback_state too often.
 483  *
 484  * Here we set ratelimit_pages to a level which ensures that when all CPUs are
 485  * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
 486  * thresholds before writeback cuts in.
 487  *
 488  * But the limit should not be set too high.  Because it also controls the
 489  * amount of memory which the balance_dirty_pages() caller has to write back.
 490  * If this is too large then the caller will block on the IO queue all the
 491  * time.  So limit it to four megabytes - the balance_dirty_pages() caller
 492  * will write six megabyte chunks, max.
 493  */
 494
 495 static void set_ratelimit(void)
 496 {
 497         ratelimit_pages = total_pages / (num_online_cpus() * 32);
 498         if (ratelimit_pages < 16)
 499                 ratelimit_pages = 16;
 500         if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
 501                 ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
 502 }
 503
 504 static int __cpuinit
 505 ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
 506 {
 507         set_ratelimit();
 508         return 0;
 509 }
 510
 511 static struct notifier_block __cpuinitdata ratelimit_nb = {
 512         .notifier_call  = ratelimit_handler,
 513         .next           = NULL,
 514 };
 515
 516 /*
 517  * If the machine has a large highmem:lowmem ratio then scale back the default
 518  * dirty memory thresholds: allowing too much dirty highmem pins an excessive
 519  * number of buffer_heads.
 520  */
 521 void __init page_writeback_init(void)
 522 {
 523         long buffer_pages = nr_free_buffer_pages();
 524         long correction;
 525
 526         total_pages = nr_free_pagecache_pages();
 527
 528         correction = (100 * 4 * buffer_pages) / total_pages;
 529
 530         if (correction < 100) {
 531                 dirty_background_ratio *= correction;
 532                 dirty_background_ratio /= 100;
 533                 vm_dirty_ratio *= correction;
 534                 vm_dirty_ratio /= 100;
 535
 536                 if (dirty_background_ratio <= 0)
 537                         dirty_background_ratio = 1;
 538                 if (vm_dirty_ratio <= 0)
 539                         vm_dirty_ratio = 1;
 540         }
 541         mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
 542         set_ratelimit();
 543         register_cpu_notifier(&ratelimit_nb);
 544 }
 545
 546 int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
 547 {
 548         int ret;
 549
 550         if (wbc->nr_to_write <= 0)
 551                 return 0;
 552         wbc->for_writepages = 1;
 553         if (mapping->a_ops->writepages)
 554                 ret = mapping->a_ops->writepages(mapping, wbc);
 555         else
 556                 ret = generic_writepages(mapping, wbc);
 557         wbc->for_writepages = 0;
 558         return ret;
 559 }
 560
 561 /**
 562  * write_one_page - write out a single page and optionally wait on I/O
 563  *
 564  * @page: the page to write
 565  * @wait: if true, wait on writeout
 566  *
 567  * The page must be locked by the caller and will be unlocked upon return.
 568  *
 569  * write_one_page() returns a negative error code if I/O failed.
 570  */
 571 int write_one_page(struct page *page, int wait)
 572 {
 573         struct address_space *mapping = page->mapping;
 574         int ret = 0;
 575         struct writeback_control wbc = {
 576                 .sync_mode = WB_SYNC_ALL,
 577                 .nr_to_write = 1,
 578         };
 579
 580         BUG_ON(!PageLocked(page));
 581
 582         if (wait)
 583                 wait_on_page_writeback(page);
 584
 585         if (clear_page_dirty_for_io(page)) {
 586                 page_cache_get(page);
 587                 ret = mapping->a_ops->writepage(page, &wbc);
 588                 if (ret == 0 && wait) {
 589                         wait_on_page_writeback(page);
 590                         if (PageError(page))
 591                                 ret = -EIO;
 592                 }
 593                 page_cache_release(page);
 594         } else {
 595                 unlock_page(page);
 596         }
 597         return ret;
 598 }
 599 EXPORT_SYMBOL(write_one_page);
 600
 601 /*
 602  * For address_spaces which do not use buffers.  Just tag the page as dirty in
 603  * its radix tree.
 604  *
 605  * This is also used when a single buffer is being dirtied: we want to set the
 606  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
 607  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
 608  *
 609  * Most callers have locked the page, which pins the address_space in memory.
 610  * But zap_pte_range() does not lock the page, however in that case the
 611  * mapping is pinned by the vma's ->vm_file reference.
 612  *
 613  * We take care to handle the case where the page was truncated from the
 614  * mapping by re-checking page_mapping() insode tree_lock.
 615  */
 616 int __set_page_dirty_nobuffers(struct page *page)
 617 {
 618         if (!TestSetPageDirty(page)) {
 619                 struct address_space *mapping = page_mapping(page);
 620                 struct address_space *mapping2;
 621
 622                 if (mapping) {
 623                         write_lock_irq(&mapping->tree_lock);
 624                         mapping2 = page_mapping(page);
 625                         if (mapping2) { /* Race with truncate? */
 626                                 BUG_ON(mapping2 != mapping);
 627                                 if (mapping_cap_account_dirty(mapping))
 628                                         __inc_zone_page_state(page,
 629                                                                 NR_FILE_DIRTY);
 630                                 radix_tree_tag_set(&mapping->page_tree,
 631                                         page_index(page), PAGECACHE_TAG_DIRTY);
 632                         }
 633                         write_unlock_irq(&mapping->tree_lock);
 634                         if (mapping->host) {
 635                                 /* !PageAnon && !swapper_space */
 636                                 __mark_inode_dirty(mapping->host,
 637                                                         I_DIRTY_PAGES);
 638                         }
 639                 }
 640                 return 1;
 641         }
 642         return 0;
 643 }
 644 EXPORT_SYMBOL(__set_page_dirty_nobuffers);
 645
 646 /*
 647  * When a writepage implementation decides that it doesn't want to write this
 648  * page for some reason, it should redirty the locked page via
 649  * redirty_page_for_writepage() and it should then unlock the page and return 0
 650  */
 651 int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
 652 {
 653         wbc->pages_skipped++;
 654         return __set_page_dirty_nobuffers(page);
 655 }
 656 EXPORT_SYMBOL(redirty_page_for_writepage);
 657
 658 /*
 659  * If the mapping doesn't provide a set_page_dirty a_op, then
 660  * just fall through and assume that it wants buffer_heads.
 661  */
 662 int fastcall set_page_dirty(struct page *page)
 663 {
 664         struct address_space *mapping = page_mapping(page);
 665
 666         if (likely(mapping)) {
 667                 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
 668                 if (spd)
 669                         return (*spd)(page);
 670                 return __set_page_dirty_buffers(page);
 671         }
 672         if (!PageDirty(page)) {
 673                 if (!TestSetPageDirty(page))
 674                         return 1;
 675         }
 676         return 0;
 677 }
 678 EXPORT_SYMBOL(set_page_dirty);
 679
 680 /*
 681  * set_page_dirty() is racy if the caller has no reference against
 682  * page->mapping->host, and if the page is unlocked.  This is because another
 683  * CPU could truncate the page off the mapping and then free the mapping.
 684  *
 685  * Usually, the page _is_ locked, or the caller is a user-space process which
 686  * holds a reference on the inode by having an open file.
 687  *
 688  * In other cases, the page should be locked before running set_page_dirty().
 689  */
 690 int set_page_dirty_lock(struct page *page)
 691 {
 692         int ret;
 693
 694         lock_page(page);
 695         ret = set_page_dirty(page);
 696         unlock_page(page);
 697         return ret;
 698 }
 699 EXPORT_SYMBOL(set_page_dirty_lock);
 700
 701 /*
 702  * Clear a page's dirty flag, while caring for dirty memory accounting.
 703  * Returns true if the page was previously dirty.
 704  */
 705 int test_clear_page_dirty(struct page *page)
 706 {
 707         struct address_space *mapping = page_mapping(page);
 708         unsigned long flags;
 709
 710         if (mapping) {
 711                 write_lock_irqsave(&mapping->tree_lock, flags);
 712                 if (TestClearPageDirty(page)) {
 713                         radix_tree_tag_clear(&mapping->page_tree,
 714                                                 page_index(page),
 715                                                 PAGECACHE_TAG_DIRTY);
 716                         write_unlock_irqrestore(&mapping->tree_lock, flags);
 717                         /*
 718                          * We can continue to use `mapping' here because the
 719                          * page is locked, which pins the address_space
 720                          */
 721                         if (mapping_cap_account_dirty(mapping)) {
 722                                 page_mkclean(page);
 723                                 dec_zone_page_state(page, NR_FILE_DIRTY);
 724                         }
 725                         return 1;
 726                 }
 727                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 728                 return 0;
 729         }
 730         return TestClearPageDirty(page);
 731 }
 732 EXPORT_SYMBOL(test_clear_page_dirty);
 733
 734 /*
 735  * Clear a page's dirty flag, while caring for dirty memory accounting.
 736  * Returns true if the page was previously dirty.
 737  *
 738  * This is for preparing to put the page under writeout.  We leave the page
 739  * tagged as dirty in the radix tree so that a concurrent write-for-sync
 740  * can discover it via a PAGECACHE_TAG_DIRTY walk.  The ->writepage
 741  * implementation will run either set_page_writeback() or set_page_dirty(),
 742  * at which stage we bring the page's dirty flag and radix-tree dirty tag
 743  * back into sync.
 744  *
 745  * This incoherency between the page's dirty flag and radix-tree tag is
 746  * unfortunate, but it only exists while the page is locked.
 747  */
 748 int clear_page_dirty_for_io(struct page *page)
 749 {
 750         struct address_space *mapping = page_mapping(page);
 751
 752         if (mapping) {
 753                 if (TestClearPageDirty(page)) {
 754                         if (mapping_cap_account_dirty(mapping)) {
 755                                 page_mkclean(page);
 756                                 dec_zone_page_state(page, NR_FILE_DIRTY);
 757                         }
 758                         return 1;
 759                 }
 760                 return 0;
 761         }
 762         return TestClearPageDirty(page);
 763 }
 764 EXPORT_SYMBOL(clear_page_dirty_for_io);
 765
 766 int test_clear_page_writeback(struct page *page)
 767 {
 768         struct address_space *mapping = page_mapping(page);
 769         int ret;
 770
 771         if (mapping) {
 772                 unsigned long flags;
 773
 774                 write_lock_irqsave(&mapping->tree_lock, flags);
 775                 ret = TestClearPageWriteback(page);
 776                 if (ret)
 777                         radix_tree_tag_clear(&mapping->page_tree,
 778                                                 page_index(page),
 779                                                 PAGECACHE_TAG_WRITEBACK);
 780                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 781         } else {
 782                 ret = TestClearPageWriteback(page);
 783         }
 784         return ret;
 785 }
 786
 787 int test_set_page_writeback(struct page *page)
 788 {
 789         struct address_space *mapping = page_mapping(page);
 790         int ret;
 791
 792         if (mapping) {
 793                 unsigned long flags;
 794
 795                 write_lock_irqsave(&mapping->tree_lock, flags);
 796                 ret = TestSetPageWriteback(page);
 797                 if (!ret)
 798                         radix_tree_tag_set(&mapping->page_tree,
 799                                                 page_index(page),
 800                                                 PAGECACHE_TAG_WRITEBACK);
 801                 if (!PageDirty(page))
 802                         radix_tree_tag_clear(&mapping->page_tree,
 803                                                 page_index(page),
 804                                                 PAGECACHE_TAG_DIRTY);
 805                 write_unlock_irqrestore(&mapping->tree_lock, flags);
 806         } else {
 807                 ret = TestSetPageWriteback(page);
 808         }
 809         return ret;
 810
 811 }
 812 EXPORT_SYMBOL(test_set_page_writeback);
 813
 814 /*
 815  * Wakes up tasks that are being throttled due to writeback congestion
 816  */
 817 void writeback_congestion_end(void)
 818 {
 819         blk_congestion_end(WRITE);
 820 }
 821 EXPORT_SYMBOL(writeback_congestion_end);
 822
 823 /*
 824  * Return true if any of the pages in the mapping are marged with the
 825  * passed tag.
 826  */
 827 int mapping_tagged(struct address_space *mapping, int tag)
 828 {
 829         unsigned long flags;
 830         int ret;
 831
 832         read_lock_irqsave(&mapping->tree_lock, flags);
 833         ret = radix_tree_tagged(&mapping->page_tree, tag);
 834         read_unlock_irqrestore(&mapping->tree_lock, flags);
 835         return ret;
 836 }
 837 EXPORT_SYMBOL(mapping_tagged);